A realistic-looking video can be synthetic, and a strange-looking video can be genuine. Reliable assessment combines source tracing, context and technical signals rather than trusting one detector score.
Begin with provenance
Find the earliest accessible publication, original creator or institution, event context and authorized versions. Compare against known records without assuming that a repost caption is original evidence. Check when and where the material supposedly originated. Synthetic content can mix authentic footage, edited speech, generative fill and unrelated sound; the answer is not always simply “real” or “fake.”
Understand detector limits
AI-media detectors classify patterns they have learned, not the historical truth of a clip. Compression, cropping, re-encoding, post-processing, language and novel generators can change error rates. Treat a detector output as one signal. A missing watermark or credential is not proof of human creation, and a visible label can itself be fabricated.
Examine harm and consent separately
A technically genuine recording may be published without consent or misleadingly excerpted. A synthetic video may be clearly disclosed and legitimately licensed. Rights violations, false endorsements and public deception require separate judgment. Do not publicly accuse a named person of fabricating content based on one automated score.
A newsroom verification process
Preserve the file and URL, record hash and first-seen time, contact the purported issuer through a known channel, compare official records and label uncertainty. For a consequential story, obtain second-person editorial review and avoid repeating unverified allegations. If the media depicts a vulnerable person, minimize redistribution of potentially harmful content.
Provenance technology
C2PA-style credentials can provide cryptographic information about a file’s declared origin and edit history when supported and correctly validated. They do not establish that every assertion shown within the media is factually true. A signed deceptive photograph is still potentially misleading.
Reader’s rule
Ask “what can I actually verify about origin, alteration and context?” rather than “what percentage does this detector think is AI?” Those are different questions with different evidence.
Sources & Further Reading
Guidance varies by jurisdiction and application. This is general education, not legal, medical or other professional advice.